Three-way experience replay for the prediction under concept drift

In the context of online time series forecasting, experience replay (ER) is a crucial technology to mitigate catastrophic forgetting arising from concept drift. ER uses historical data sampled from a buffer to update model parameters along with current data incrementally. However, existing ER methods overlook buffer management, which is certainly important due to the space constraints and dynamic data distributions. To fill this gap, we propose the three-way experience replay (TWER). In contrast to existing ER methods, TWER is a comprehensive buffer management mechanism that involves admission, sampling, and eviction to improve the accuracy of online prediction. To capture samples indicative of concept drift, we design the admission mechanism based on three-way decision (TWD) rules deduced from historical and future data. By integrating Bayesian decision theory and rough set theory, TWD can provide the option of deferment action, enabling a more informed judgment on concept drift. To preserve the historical data with rare patterns, we enhance the sampling and eviction mechanisms by introducing three-way clustering (TWC), which can identify fringe samples typically missed by existing methods. Leveraging these mechanisms, TWER can adaptively manage data within the buffer, thus enhancing the prediction under concept drift. Empirical results on five real-world datasets show that TWER reduces online prediction error by more than 12% in terms of MSE compared with the existing ER methods such as DER++, GDumb, and memory select.